Add user provided JSON operator list build

Allow ACL to be built via a user provided JSON file containing operators, data types and data layouts.
Modify TFLite file to JSON file script to output data layouts.
Fix build issue with "fat_binary" and "high_priority" options.

Resolves: COMPMID-4697, COMPMID-4837

Signed-off-by: Freddie Liardet <frederick.liardet@arm.com>
Change-Id: I08d494151c98f804325707ffd922ffe216813023
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6427
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>

26 files changed, 399 insertions, 389 deletions

diff --git a/SConscript b/SConscript
index 6672caee9f..bcb93fde62 100644
--- a/SConscript
+++ b/SConscript
@@ -157,31 +157,28 @@ def create_version_file(target, source, env):
         fd.write(build_info)
 
 
-def get_attrs_list(arch, estate, data_types, data_layouts):
+def get_attrs_list(env, data_types, data_layouts):
     attrs = []
 
     # Manage data-types
-    if any(i in data_types for i in ['all']):
+    if 'all' in data_types:
         attrs += ['fp16', 'fp32', 'integer', 'qasymm8', 'qasymm8_signed', 'qsymm16']
     else:
-        if any(i in data_types for i in ['fp16']): attrs += ['fp16']
-        if any(i in data_types for i in ['fp32']): attrs += ['fp32']
-        if any(i in data_types for i in ['integer']): attrs += ['integer']
-        if any(i in data_types for i in ['qasymm8']): attrs += ['qasymm8']
-        if any(i in data_types for i in ['qasymm8_signed']): attrs += ['qasymm8_signed']
-        if any(i in data_types for i in ['qsymm16']): attrs += ['qsymm16']
-
+        if 'fp16' in data_types: attrs += ['fp16']
+        if 'fp32' in data_types: attrs += ['fp32']
+        if 'integer' in data_types: attrs += ['integer']
+        if 'qasymm8' in data_types: attrs += ['qasymm8']
+        if 'qasymm8_signed' in data_types: attrs += ['qasymm8_signed']
+        if 'qsymm16' in data_types: attrs += ['qsymm16']
     # Manage data-layouts
-    if any(i in data_layouts for i in ['all']):
+    if 'all' in data_layouts:
         attrs += ['nhwc', 'nchw']
     else:
-        if any(i in data_layouts for i in ['nhwc']): attrs += ['nhwc']
-        if any(i in data_layouts for i in ['nchw']): attrs += ['nchw']
+        if 'nhwc' in data_layouts: attrs += ['nhwc']
+        if 'nchw' in data_layouts: attrs += ['nchw']
 
     # Manage execution state
-    estate_attr = 'estate32' if (estate == 'auto' and 'v7a' in arch) or '32' in estate else 'estate64'
-    attrs += [ estate_attr ]
-
+    attrs += ['estate32' if (env['estate'] == 'auto' and 'v7a' in env['arch']) or '32' in env['estate'] else 'estate64']
     return attrs
 
 
@@ -237,6 +234,27 @@ def resolve_operator_dependencies(filelist, operators, backend=''):
 
     return resolved_operators
 
+def read_build_config_json(build_config):
+    build_config_contents = {}
+    custom_operators = []
+    custom_types = []
+    custom_layouts = []
+    if os.path.isfile(build_config):
+        with open(build_config) as f:
+            try:
+                build_config_contents = json.load(f)
+            except:
+                print("Warning: Build configuration file is of invalid JSON format!")
+    else:
+        try:
+            build_config_contents = json.loads(build_config)
+        except:
+            print("Warning: Build configuration string is of invalid JSON format!")
+    if build_config_contents:
+        custom_operators = build_config_contents.get("operators", [])
+        custom_types = build_config_contents.get("data_types", [])
+        custom_layouts = build_config_contents.get("data_layouts", [])
+    return custom_operators, custom_types, custom_layouts
 
 arm_compute_env = env.Clone()
 version_file = arm_compute_env.Command("src/core/arm_compute_version.embed", "", action=create_version_file)
@@ -427,30 +445,25 @@ graph_files = Glob('src/graph/*.cpp')
 graph_files += Glob('src/graph/*/*.cpp')
 
 # Specify user-defined priority operators
-use_priority_ops = env['high_priority']
-priority_operators = filelist['high_priority']
-if env['build_config'] != "":
-    build_config = env['build_config']
-    build_config_contents = {}
-    if os.path.isfile(build_config):
-        with open(build_config) as f:
-            try:
-                build_config_contents = json.load(f)
-            except:
-                print("Warning: Build configuration file is of invalid JSON format!")
-    else:
-        try:
-            build_config_contents = json.loads(build_config)
-        except:
-            print("Warning: Build configuration string is of invalid JSON format!")
-    if build_config_contents:
-        priority_operators = build_config_contents.get("operators", [])
+custom_operators = []
+custom_types = []
+custom_layouts = []
+
+use_custom_ops = env['high_priority'] or env['build_config'];
+
+if env['high_priority']:
+    custom_operators = filelist['high_priority']
+    custom_types = ['all']
+    custom_layouts = ['all']
+
+if env['build_config']:
+    custom_operators, custom_types, custom_layouts = read_build_config_json(env['build_config'])
 
 if env['opencl']:
     lib_files += filelist['c_api']['gpu']
     lib_files += filelist['gpu']['common']
 
-    cl_operators = priority_operators if use_priority_ops else filelist['gpu']['operators'].keys()
+    cl_operators = custom_operators if use_custom_ops else filelist['gpu']['operators'].keys()
     cl_ops_to_build = resolve_operator_dependencies(filelist, cl_operators, 'gpu')
     lib_files += get_operator_backend_files(filelist, cl_ops_to_build, 'gpu')['common']
 
@@ -475,11 +488,15 @@ if env['neon']:
     if 'sve' not in env['arch'] or env['fat_binary']: simd += ['neon']
 
     # Get attributes
-    attrs = get_attrs_list(env['arch'], env['estate'], env['data_type_support'], env['data_layout_support'])
+    if(use_custom_ops):
+        attrs = get_attrs_list(env, custom_types, custom_layouts)
+    else:
+        attrs = get_attrs_list(env, env['data_type_support'], env['data_layout_support'])
 
     # Setup data-type and data-layout files to include
-    cpu_operators = priority_operators if use_priority_ops else filelist['cpu']['operators'].keys()
-    cpu_ops_to_build = resolve_operator_dependencies(filelist, filelist['cpu']['operators'], 'cpu')
+    cpu_operators = custom_operators if use_custom_ops else filelist['cpu']['operators'].keys()
+    cpu_ops_to_build = resolve_operator_dependencies(filelist, cpu_operators, 'cpu')
+
     cpu_files = get_operator_backend_files(filelist, cpu_ops_to_build, 'cpu', simd, attrs)
     lib_files += cpu_files.get('common', [])
     lib_files += cpu_files.get('neon', [])
@@ -488,8 +505,8 @@ if env['neon']:
     graph_files += Glob('src/graph/backends/NEON/*.cpp')
 
 # Restrict from building graph API if a reduced operator list has been provided
-if use_priority_ops:
-    print("Graph library requires all operators to be built")
+if use_custom_ops:
+    print("WARNING: Graph library requires all operators to be built")
     graph_files = []
 
 # Build bootcode in case of bare-metal
diff --git a/SConstruct b/SConstruct
index 7591075cd1..400228c71a 100644
--- a/SConstruct
+++ b/SConstruct
@@ -41,6 +41,50 @@ def version_at_least(version, required):
 
     return True
 
+def read_build_config_json(build_config):
+    build_config_contents = {}
+    custom_types = []
+    custom_layouts = []
+    if os.path.isfile(build_config):
+        with open(build_config) as f:
+            try:
+                build_config_contents = json.load(f)
+            except:
+                print("Warning: Build configuration file is of invalid JSON format!")
+    else:
+        try:
+            build_config_contents = json.loads(build_config)
+        except:
+            print("Warning: Build configuration string is of invalid JSON format!")
+    if build_config_contents:
+        custom_types = build_config_contents.get("data_types", [])
+        custom_layouts = build_config_contents.get("data_layouts", [])
+    return custom_types, custom_layouts
+
+def update_data_type_layout_flags(env, data_types, data_layouts):
+    # Manage data-types
+    if any(i in data_types for i in ['all', 'fp16']):
+        env.Append(CXXFLAGS = ['-DENABLE_FP16_KERNELS'])
+    if any(i in data_types for i in ['all', 'fp32']):
+        env.Append(CXXFLAGS = ['-DENABLE_FP32_KERNELS'])
+    if any(i in data_types for i in ['all', 'qasymm8']):
+        env.Append(CXXFLAGS = ['-DENABLE_QASYMM8_KERNELS'])
+    if any(i in data_types for i in ['all', 'qasymm8_signed']):
+        env.Append(CXXFLAGS = ['-DENABLE_QASYMM8_SIGNED_KERNELS'])
+    if any(i in data_types for i in ['all', 'qsymm16']):
+        env.Append(CXXFLAGS = ['-DENABLE_QSYMM16_KERNELS'])
+    if any(i in data_types for i in ['all', 'integer']):
+        env.Append(CXXFLAGS = ['-DENABLE_INTEGER_KERNELS'])
+
+    # Manage data-layouts
+    if any(i in data_layouts for i in ['all', 'nhwc']):
+        env.Append(CXXFLAGS = ['-DENABLE_NHWC_KERNELS'])
+    if any(i in data_layouts for i in ['all', 'nchw']):
+        env.Append(CXXFLAGS = ['-DENABLE_NCHW_KERNELS'])
+
+    return env
+
+
 vars = Variables("scons")
 vars.AddVariables(
     BoolVariable("debug", "Debug", False),
@@ -327,25 +371,20 @@ if env['high_priority'] and env['build_config']:
 if not env['high_priority'] and not env['build_config']:
     env.Append(CPPDEFINES = ['ARM_COMPUTE_GRAPH_ENABLED'])
 
-if env['data_type_support']:
-    if any(i in env['data_type_support'] for i in ['all', 'fp16']):
-        env.Append(CXXFLAGS = ['-DENABLE_FP16_KERNELS'])
-    if any(i in env['data_type_support'] for i in ['all', 'fp32']):
-        env.Append(CXXFLAGS = ['-DENABLE_FP32_KERNELS'])
-    if any(i in env['data_type_support'] for i in ['all', 'qasymm8']):
-        env.Append(CXXFLAGS = ['-DENABLE_QASYMM8_KERNELS'])
-    if any(i in env['data_type_support'] for i in ['all', 'qasymm8_signed']):
-        env.Append(CXXFLAGS = ['-DENABLE_QASYMM8_SIGNED_KERNELS'])
-    if any(i in env['data_type_support'] for i in ['all', 'qsymm16']):
-        env.Append(CXXFLAGS = ['-DENABLE_QSYMM16_KERNELS'])
-    if any(i in env['data_type_support'] for i in ['all', 'integer']):
-        env.Append(CXXFLAGS = ['-DENABLE_INTEGER_KERNELS'])
+data_types = []
+data_layouts = []
 
-if env['data_layout_support']:
-    if any(i in env['data_layout_support'] for i in ['all', 'nhwc']):
-        env.Append(CXXFLAGS = ['-DENABLE_NHWC_KERNELS'])
-    if any(i in env['data_layout_support'] for i in ['all', 'nchw']):
-        env.Append(CXXFLAGS = ['-DENABLE_NCHW_KERNELS'])
+# Set correct data types / layouts to build
+if env['high_priority']:
+    data_types = ['all']
+    data_layouts = ['all']
+elif env['build_config']:
+    data_types, data_layouts = read_build_config_json(env['build_config'])
+else:
+    data_types = env['data_type_support']
+    data_layouts = env['data_layout_support']
+
+env = update_data_type_layout_flags(env, data_types, data_layouts)
 
 if env['standalone']:
     env.Append(CXXFLAGS = ['-fPIC'])
@@ -417,6 +456,10 @@ Export('version_at_least')
 
 SConscript('./SConscript', variant_dir=build_path, duplicate=0)
 
+if env['examples'] and (env['build_config'] or env['high_priority']):
+    print("WARNING: Building examples for selected operators not supported. Use examples=0")
+    Return()
+
 if env['examples'] and env['exceptions']:
     if env['os'] == 'bare_metal' and env['arch'] == 'armv7a':
         print("WARNING: Building examples for bare metal and armv7a is not supported. Use examples=0")
@@ -424,6 +467,9 @@ if env['examples'] and env['exceptions']:
     SConscript('./examples/SConscript', variant_dir='%s/examples' % build_path, duplicate=0)
 
 if env['exceptions']:
+    if env['build_config'] or env['high_priority']:
+        print("WARNING: Building tests for selected operators not supported")
+        Return()
     if env['os'] == 'bare_metal' and env['arch'] == 'armv7a':
         print("WARNING: Building tests for bare metal and armv7a is not supported")
         Return()
diff --git a/filelist.json b/filelist.json
index e52b7c824c..bcc7ecb37a 100644
--- a/filelist.json
+++ b/filelist.json
@@ -845,21 +845,21 @@
           "common": [
             "src/cpu/operators/CpuActivation.cpp",
             "src/cpu/kernels/CpuActivationKernel.cpp",
-            "src/runtime/NEON/functions/NEActivationLayer.cpp"
+            "src/runtime/NEON/functions/NEActivationLayer.cpp",
+            "src/cpu/kernels/activation/neon/qasymm8.cpp",
+            "src/cpu/kernels/activation/neon/qasymm8_signed.cpp",
+            "src/cpu/kernels/activation/neon/qsymm16.cpp"
           ],
           "neon": {
             "fp16": [ "src/cpu/kernels/activation/neon/fp16.cpp" ],
-            "fp32": [ "src/cpu/kernels/activation/neon/fp32.cpp" ],
-            "qasymm8": [ "src/cpu/kernels/activation/neon/qasymm8.cpp" ],
-            "qasymm8_signed": [ "src/cpu/kernels/activation/neon/qasymm8_signed.cpp" ],
-            "qsymm16": [ "src/cpu/kernels/activation/neon/qsymm16.cpp" ]
+            "fp32": [ "src/cpu/kernels/activation/neon/fp32.cpp" ]
           },
           "sve": {
             "fp16": [ "src/cpu/kernels/activation/sve/fp16.cpp" ],
             "fp32": [ "src/cpu/kernels/activation/sve/fp32.cpp" ],
-            "qasymm8": [ "src/cpu/kernels/activation/neon/qasymm8.cpp", "src/cpu/kernels/activation/sve/qasymm8.cpp" ],
-            "qasymm8_signed": [ "src/cpu/kernels/activation/neon/qasymm8_signed.cpp", "src/cpu/kernels/activation/sve/qasymm8_signed.cpp" ],
-            "qsymm16": [ "src/cpu/kernels/activation/neon/qsymm16.cpp", "src/cpu/kernels/activation/sve/qsymm16.cpp" ]
+            "qasymm8": [ "src/cpu/kernels/activation/sve/qasymm8.cpp" ],
+            "qasymm8_signed": [ "src/cpu/kernels/activation/sve/qasymm8_signed.cpp" ],
+            "qsymm16": [ "src/cpu/kernels/activation/sve/qsymm16.cpp" ]
           }
         }
       },
@@ -874,18 +874,16 @@
           "common": [
             "src/cpu/operators/CpuAdd.cpp",
             "src/cpu/kernels/CpuAddKernel.cpp",
-            "src/runtime/NEON/functions/NEArithmeticAddition.cpp"
+            "src/runtime/NEON/functions/NEArithmeticAddition.cpp",
+            "src/cpu/kernels/add/neon/qasymm8.cpp",
+            "src/cpu/kernels/add/neon/qasymm8_signed.cpp",
+            "src/cpu/kernels/add/neon/qsymm16.cpp"
           ],
-          "neon": {
-            "qasymm8": [ "src/cpu/kernels/add/neon/qasymm8.cpp" ],
-            "qasymm8_signed": [ "src/cpu/kernels/add/neon/qasymm8_signed.cpp" ],
-            "qsymm16": [ "src/cpu/kernels/add/neon/qsymm16.cpp" ]
-          },
           "sve": {
             "common": [ "src/cpu/kernels/add/sve/impl.cpp" ],
-            "qasymm8": [ "src/cpu/kernels/add/neon/qasymm8.cpp", "src/cpu/kernels/add/sve/qasymm8.cpp" ],
-            "qasymm8_signed": [ "src/cpu/kernels/add/neon/qasymm8_signed.cpp", "src/cpu/kernels/add/sve/qasymm8_signed.cpp" ],
-            "qsymm16": [ "src/cpu/kernels/add/neon/qsymm16.cpp", "src/cpu/kernels/add/sve/qsymm16.cpp" ]
+            "qasymm8": [ "src/cpu/kernels/add/sve/qasymm8.cpp" ],
+            "qasymm8_signed": [ "src/cpu/kernels/add/sve/qasymm8_signed.cpp" ],
+            "qsymm16": [ "src/cpu/kernels/add/sve/qsymm16.cpp" ]
           }
         }
       },
@@ -1103,68 +1101,62 @@
             "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp",
             "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp",
             "src/core/NEON/kernels/arm_conv/depthwise/interleaves/8b_mla.cpp",
-            "src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp"
+            "src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_generic_output9_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp",
+            "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp"
           ],
-          "neon": {
-            "estate64": [
-              "src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp"
-            ]
-          },
           "sve": {
             "common": [
-              "src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp",
               "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_8b_mla.cpp",
               "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp",
               "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp",
@@ -1209,57 +1201,7 @@
               "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
               "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
               "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp"
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp"
             ]
           }
         }
@@ -1316,6 +1258,7 @@
         }
       },
       "FFT1D": {
+        "deps": [ "Reduction" ],
         "files": {
           "common": [
             "src/core/NEON/kernels/NEFFTDigitReverseKernel.cpp",
@@ -1385,6 +1328,7 @@
         }
       },
       "Gemm": {
+        "deps": [ "Quantize", "Add"],
         "files": {
           "common": [
             "src/cpu/kernels/CpuConvertQuantizedSignednessKernel.cpp",
@@ -1422,7 +1366,61 @@
             "src/core/NEON/kernels/arm_gemm/transform.cpp",
             "src/runtime/NEON/functions/NEGEMM.cpp",
             "src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp",
-            "src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp"
+            "src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/generic.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/a55.cpp",
+            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/generic.cpp"
           ],
           "neon": {
             "estate32": [
@@ -1431,68 +1429,14 @@
               "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/generic.cpp"
             ],
             "estate64": [
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/x1.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemv_pretransposed/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/generic.cpp"
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemv_pretransposed/generic.cpp"
             ]
           },
           "sve": {
@@ -1536,69 +1480,7 @@
               "src/core/NEON/kernels/arm_gemm/kernels/sve_smallK_hybrid_s8s32_dot_8x1VL/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/sve_smallK_hybrid_u8u32_dot_8x1VL/generic.cpp",
               "src/core/NEON/kernels/arm_gemm/mergeresults-sve.cpp",
-              "src/core/NEON/kernels/arm_gemm/transform-sve.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/x1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemv_pretransposed/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/generic.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/a55.cpp",
-              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/generic.cpp"
+              "src/core/NEON/kernels/arm_gemm/transform-sve.cpp"
             ]
           }
         }
@@ -1735,38 +1617,34 @@
             "src/core/NEON/kernels/arm_conv/pooling/pooling_s8q.cpp",
             "src/core/NEON/kernels/arm_conv/pooling/pooling_u8.cpp",
             "src/core/NEON/kernels/arm_conv/pooling/pooling_u8q.cpp",
-            "src/runtime/NEON/functions/NEPoolingLayer.cpp"
+            "src/runtime/NEON/functions/NEPoolingLayer.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_generic_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_generic_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_generic_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_avg_generic_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_generic_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_avg_generic_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_avg_generic_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_generic_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
+            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp",
+            "src/cpu/kernels/pool2d/neon/qasymm8.cpp",
+            "src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp"
           ],
           "neon": {
             "nchw": [ "src/cpu/kernels/pool2d/neon/nchw/all.cpp" ],
             "fp16": [ "src/cpu/kernels/pool2d/neon/fp16.cpp" ],
-            "fp32": [ "src/cpu/kernels/pool2d/neon/fp32.cpp" ],
-            "qasymm8": [ "src/cpu/kernels/pool2d/neon/qasymm8.cpp" ],
-            "qasymm8_signed": [ "src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp" ],
-            "estate64": [
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp"
-            ]
+            "fp32": [ "src/cpu/kernels/pool2d/neon/fp32.cpp" ]
           },
           "sve": {
-            "qasymm8": [ "src/cpu/kernels/pool2d/neon/qasymm8.cpp" ],
-            "qasymm8_signed": [ "src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp" ],
             "common": [
               "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
               "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
@@ -1785,25 +1663,7 @@
               "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
               "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_max_generic_depthfirst/generic.cpp",
               "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_max_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
-              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp"
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8q_nhwc_max_generic_depthfirst/generic.cpp"
             ]
           }
         }
@@ -2002,18 +1862,11 @@
           "common": [
             "src/cpu/operators/CpuSub.cpp",
             "src/cpu/kernels/CpuSubKernel.cpp",
-            "src/runtime/NEON/functions/NEArithmeticSubtraction.cpp"
-          ],
-          "sve": {
-            "qasymm8": [ "src/cpu/kernels/sub/neon/qasymm8.cpp" ],
-            "qasymm8_signed": [ "src/cpu/kernels/sub/neon/qasymm8_signed.cpp" ],
-            "qsymm16": [ "src/cpu/kernels/sub/neon/qsymm16.cpp" ]
-          },
-          "neon": {
-            "qasymm8": [ "src/cpu/kernels/sub/neon/qasymm8.cpp" ],
-            "qasymm8_signed": [ "src/cpu/kernels/sub/neon/qasymm8_signed.cpp" ],
-            "qsymm16": [ "src/cpu/kernels/sub/neon/qsymm16.cpp" ]
-          }
+            "src/runtime/NEON/functions/NEArithmeticSubtraction.cpp",
+            "src/cpu/kernels/sub/neon/qasymm8.cpp",
+            "src/cpu/kernels/sub/neon/qasymm8_signed.cpp",
+            "src/cpu/kernels/sub/neon/qsymm16.cpp"
+          ]
         }
       },
       "Tile": {
diff --git a/python/scripts/report-model-ops/report_model_ops.py b/python/scripts/report-model-ops/report_model_ops.py
index 3888b801e6..1549005da5 100644..100755
--- a/python/scripts/report-model-ops/report_model_ops.py
+++ b/python/scripts/report-model-ops/report_model_ops.py
@@ -31,20 +31,20 @@ import tflite
 sys.path.append(os.path.dirname(os.path.abspath(__file__)) + "/../")
 
 from utils.model_identification import identify_model_type
-from utils.tflite_helpers import tflite_op2acl, tflite_typecode2name
+from utils.tflite_helpers import tflite_op2acl, tflite_typecode2name, tflite_typecode2aclname
 
 SUPPORTED_MODEL_TYPES = ["tflite"]
 logger = logging.getLogger("report_model_ops")
 
 
-def get_ops_from_tflite_graph(model):
+def get_ops_types_from_tflite_graph(model):
     """
-    Helper function that extract operator related meta-data from a TfLite model
+    Helper function that extract operator related meta-data from a TFLite model
 
     Parameters
         ----------
     model: str
-        Respective TfLite model to analyse
+        Respective TFLite model to analyse
 
     Returns
     ----------
@@ -52,7 +52,7 @@ def get_ops_from_tflite_graph(model):
         A tuple with the sets of unique operator types and data-types that are present in the model
     """
 
-    logger.debug(f"Analysing TfLite mode '{model}'!")
+    logger.debug(f"Analysing TFLite mode '{model}'!")
 
     with open(model, "rb") as f:
         buf = f.read()
@@ -63,11 +63,16 @@ def get_ops_from_tflite_graph(model):
     unique_ops = {tflite.opcode2name(model.OperatorCodes(op_id).BuiltinCode()) for op_id in range(0, nr_unique_ops)}
 
     # Extract IO data-types
-    data_types = set()
+    supported_data_types = set()
+    unsupported_data_types = set()
     for subgraph_id in range(0, model.SubgraphsLength()):
         subgraph = model.Subgraphs(subgraph_id)
         for tensor_id in range(0, subgraph.TensorsLength()):
-            data_types.add(tflite_typecode2name(subgraph.Tensors(tensor_id).Type()))
+            try:
+                supported_data_types.add(tflite_typecode2aclname(subgraph.Tensors(tensor_id).Type()))
+            except ValueError:
+                unsupported_data_types.add(tflite_typecode2name(subgraph.Tensors(tensor_id).Type()))
+                logger.warning(f"Data type {tflite_typecode2name(subgraph.Tensors(tensor_id).Type())} is not supported by ComputeLibrary")
 
     # Perform mapping between TfLite ops to ComputeLibrary ones
     supported_ops = set()
@@ -75,17 +80,17 @@ def get_ops_from_tflite_graph(model):
     for top in unique_ops:
         try:
             supported_ops.add(tflite_op2acl(top))
-        except:
+        except ValueError:
             unsupported_ops.add(top)
-            logger.warning(f"Operator {top} has not ComputeLibrary mapping")
+            logger.warning(f"Operator {top} does not have ComputeLibrary mapping")
 
-    return (supported_ops, unsupported_ops, data_types)
+    return (supported_ops, unsupported_ops, supported_data_types, unsupported_data_types)
 
 
 def extract_model_meta(model, model_type):
     """
     Function that calls the appropriate model parser to extract model related meta-data
-    Supported parsers: TfLite
+    Supported parsers: TFLite
 
     Parameters
         ----------
@@ -101,13 +106,13 @@ def extract_model_meta(model, model_type):
     """
 
     if model_type == "tflite":
-        return get_ops_from_tflite_graph(model)
+        return get_ops_types_from_tflite_graph(model)
     else:
         logger.warning(f"Model type '{model_type}' is unsupported!")
         return ()
 
 
-def generate_build_config(ops, data_types):
+def generate_build_config(ops, data_types, data_layouts):
     """
     Function that generates a compatible ComputeLibrary operator-based build configuration
 
@@ -117,6 +122,8 @@ def generate_build_config(ops, data_types):
         Set with the operators to add in the build configuration
     data_types:
         Set with the data types to add in the build configuration
+    data_layouts:
+        Set with the data layouts to add in the build configuration
 
     Returns
     ----------
@@ -126,6 +133,7 @@ def generate_build_config(ops, data_types):
     config_data = {}
     config_data["operators"] = list(ops)
     config_data["data_types"] = list(data_types)
+    config_data["data_layouts"] = list(data_layouts)
 
     return config_data
 
@@ -134,7 +142,7 @@ if __name__ == "__main__":
     parser = ArgumentParser(
         description="""Report map of operations in a list of models.
             The script consumes deep learning models and reports the type of operations and data-types used
-            Supported model types: TfLite """
+            Supported model types: TFLite """
     )
 
     parser.add_argument(
@@ -163,26 +171,35 @@ if __name__ == "__main__":
     # Extract operator mapping
     final_supported_ops = set()
     final_unsupported_ops = set()
-    final_dts = set()
+    final_supported_dts = set()
+    final_unsupported_dts = set()
+    final_layouts = {"nhwc"} # Data layout for TFLite is always NHWC
     for model in args.models:
         logger.debug(f"Starting analyzing {model} model")
 
         model_type = identify_model_type(model)
-        supported_model_ops, unsupported_mode_ops, model_dts = extract_model_meta(model, model_type)
+        supported_model_ops, unsupported_mode_ops, supported_model_dts, unsupported_model_dts = extract_model_meta(model, model_type)
         final_supported_ops.update(supported_model_ops)
         final_unsupported_ops.update(unsupported_mode_ops)
-        final_dts.update(model_dts)
+        final_supported_dts.update(supported_model_dts)
+        final_unsupported_dts.update(unsupported_model_dts)
 
     logger.info("=== Supported Operators")
     logger.info(final_supported_ops)
-    logger.info("=== Unsupported Operators")
-    logger.info(final_unsupported_ops)
+    if(len(final_unsupported_ops)):
+        logger.info("=== Unsupported Operators")
+        logger.info(final_unsupported_ops)
     logger.info("=== Data Types")
-    logger.info(final_dts)
-
-    # Generate json file
+    logger.info(final_supported_dts)
+    if(len(final_unsupported_dts)):
+        logger.info("=== Unsupported Data Types")
+        logger.info(final_unsupported_dts)
+    logger.info("=== Data Layouts")
+    logger.info(final_layouts)
+
+    # Generate JSON file
     if args.config:
         logger.debug("Generating JSON build configuration file")
-        config_data = generate_build_config(final_supported_ops, final_dts)
+        config_data = generate_build_config(final_supported_ops, final_supported_dts, final_layouts)
         with open(args.config, "w") as f:
             json.dump(config_data, f)
diff --git a/python/scripts/utils/model_identification.py b/python/scripts/utils/model_identification.py
index 43e7d20f61..84a6e1a097 100644
--- a/python/scripts/utils/model_identification.py
+++ b/python/scripts/utils/model_identification.py
@@ -24,7 +24,7 @@ import os
 
 
 def is_tflite_model(model_path):
-    """Check if a model is of TfLite type
+    """Check if a model is of TFLite type
 
     Parameters:
     ----------
@@ -34,7 +34,7 @@ def is_tflite_model(model_path):
     Returns
     ----------
     bool:
-        True if given path is a valid TfLite model
+        True if given path is a valid TFLite model
     """
 
     try:
diff --git a/python/scripts/utils/tflite_helpers.py b/python/scripts/utils/tflite_helpers.py
index 8f8d422743..c2aeaac6a7 100644
--- a/python/scripts/utils/tflite_helpers.py
+++ b/python/scripts/utils/tflite_helpers.py
@@ -20,6 +20,19 @@
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 # SOFTWARE.
 
+_TFLITE_TYPECODE2ACLNAME = {
+    0: "fp32", # Float32
+    1: "fp16", # Float16
+    2: "integer", # Int32
+    3: "qasymm8", # Uint8
+    # 4: "Unsupported", # Int64
+    # 5: "Unsupported", # String
+    6: "integer", # Bool
+    7: "qsymm16", # Int16
+    # 8: "Unsupported", # Complex64
+    9: "qasymm8_signed", # Int8
+}
+
 _TFLITE_TYPECODE2NAME = {
     0: "Float32",
     1: "Float16",
@@ -182,13 +195,36 @@ _TFLITE_TO_ACL = {
 }
 
 
+def tflite_typecode2aclname(toc):
+    """Stringify TFLite data-type opcodes to ACL versions
+
+    Parameters:
+    ----------
+    toc: int
+        TFLite type opcode
+
+    Returns
+    ----------
+    str
+        Stringified opcode
+
+    Raises
+    ------
+    ValueError
+        If opcode does not exist in the map
+    """
+    if toc in _TFLITE_TYPECODE2ACLNAME:
+        return _TFLITE_TYPECODE2ACLNAME[toc]
+    else:
+        raise ValueError("Unknown ACL typecode %d" % toc)
+
 def tflite_typecode2name(toc):
-    """Stringify TfLite data-type opcodes
+    """Stringify TFLite data-type opcodes
 
     Parameters:
     ----------
     toc: int
-        TfLite type opcode
+        TFLite type opcode
 
     Returns
     ----------
@@ -207,12 +243,12 @@ def tflite_typecode2name(toc):
 
 
 def tflite_op2acl(top):
-    """Map TfLite operators to ComputeLibrary ones
+    """Map TFLite operators to ComputeLibrary ones
 
     Parameters:
     ----------
     top: str
-        TfLite operator name
+        TFLite operator name
 
     Returns
     ----------
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp
index 7a26ba4230..5107ddacbc 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp
@@ -72,6 +72,7 @@ namespace
           );
   }
 
+#if defined(__aarch64__)
   unsigned int not_preferred(const DepthwiseArgs &, const Nothing &)
   {
     return std::numeric_limits<unsigned int>::max();
@@ -81,6 +82,7 @@ namespace
   {
     return args.channel_multiplier > 1 ? 0 : std::numeric_limits<unsigned int>::max();
   }
+#endif // defined(__aarch64__)
 }
 
 static const DepthwiseImplementation<float, float> depthwise_fp32_methods[] = {
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp
index 1c4c7576f5..46a31185d7 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp
@@ -62,11 +62,13 @@ namespace depthwise {
 
 namespace
 {
+#if defined(__aarch64__)
 bool qp_weights_are_symmetric(const DepthwiseArgs &, const void *_qp)
 {
   const auto qp = static_cast<const arm_gemm::Requantize32 *>(_qp);
   return qp->b_offset == 0;
 }
+#endif // defined(__aarch64__)
 }
 
 static const DepthwiseImplementation<int8_t, int8_t, int8_t, Requantize32> depthwise_s8q_methods[] = {
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp
index e8e817e9cc..c0b87ada75 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 #include <cstddef>
 #include <cstdint>
 
@@ -377,3 +378,4 @@ void a64_fp32_nhwc_generic_output9_mla_depthfirst_impl(
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp
index 5e334ec7b8..04a7abd3bd 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 #include <cstddef>
 #include <cstdint>
 
@@ -530,3 +531,4 @@ void a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst_imp
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp
index 6e9e97fa29..67fc09b2ee 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp
@@ -22,6 +22,8 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
+
 #include <cstddef>
 #include <cstdint>
 
@@ -914,3 +916,4 @@ void a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst_imp
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
index c93037d183..46210e2964 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 #include <cstddef>
 #include <cstdint>
 
@@ -849,3 +850,4 @@ void a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst_im
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp
index ad5545a304..78f748ad58 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include "arm_gemm.hpp"
 #include <cstddef>
@@ -622,3 +623,4 @@ void a64_s8q_nhwc_generic_output9_mla_depthfirst_impl(
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp
index 2fb6d3538f..cbe3d2cd1c 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include "arm_gemm.hpp"
 #include <cstddef>
@@ -525,3 +526,4 @@ void a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst_impl
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
index 95ad78cf6c..b198eff6ac 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include "arm_gemm.hpp"
 #include <cstddef>
@@ -660,3 +661,4 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
index c0acd8805e..bbfa9f439f 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include "arm_gemm.hpp"
 #include <cstddef>
@@ -1482,3 +1483,4 @@ void a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst_imp
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp
index 42d9b2f408..9cebfe8f03 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include "arm_gemm.hpp"
 #include <cstddef>
@@ -622,3 +623,4 @@ void a64_u8q_nhwc_generic_output9_mla_depthfirst_impl(
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp
index 2106cf7086..057b1ef492 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include "arm_gemm.hpp"
 #include <cstddef>
@@ -525,3 +526,4 @@ void a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst_impl
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
index 8bcd682e3c..40242e9718 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include "arm_gemm.hpp"
 #include <cstddef>
@@ -660,3 +661,4 @@ void a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
index ada1818eba..e896304c59 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include "arm_gemm.hpp"
 #include <cstddef>
@@ -1482,3 +1483,4 @@ void a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst_imp
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp
index 1633639ad5..08a2b7a98e 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include "arm_gemm.hpp"
 #include <cstddef>
@@ -622,3 +623,4 @@ void a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst_impl(
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
index 152999dd1a..09b274056f 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include "arm_gemm.hpp"
 #include <cstddef>
@@ -1482,3 +1483,4 @@ void a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst
 
 }  // namespace depthwise
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp
index ff8d7d8ba1..71a8c7496a 100644
--- a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 #include <algorithm>
 #include <cstddef>
 #include <cstdint>
@@ -249,3 +250,4 @@ void a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst_impl(
 
 }  // namespace pooling
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp
index ea7e2195d1..a924c9a7a6 100644
--- a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include <cstddef>
 #include <cstdint>
@@ -172,3 +173,4 @@ void a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst_impl(
 
 }  // namespace pooling
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp
index 298db96861..e344e14e34 100644
--- a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include <cstddef>
 #include <cstdint>
@@ -172,3 +173,4 @@ void a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst_impl(
 
 }  // namespace pooling
 }  // namespace arm_conv
+#endif // defined(__aarch64__)
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp
index 02c43ccaba..9d379d183e 100644
--- a/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 
+#if defined(__aarch64__)
 
 #include <cstddef>
 #include <cstdint>
@@ -172,3 +173,4 @@ void a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst_impl(
 
 }  // namespace pooling
 }  // namespace arm_conv
+#endif // defined(__aarch64__)